Broken syntax in oral speech, but not a reparandum

[shvedova]

Hello, I am working with transcripts of Ukrainian parliamentary sessions, which are written down verbatim. There are some grammatical errors due to the contamination of two different structures, which happens in oral speech.
What's the best way to annotate this?

Here's an example:
Моє — my.NOM.N.SG
переконання — conviction.NOM.N.SG
:
настав — has come.M.SG.PST
час — time.NOM.M.SG
відмовитися — to refuse.INF
від — from
дій — actions.GEN.PL
,
надиктованих — dictated.PTCP.PASS.GEN.PL
методів — methods.GEN.PL
політичної — political.GEN.F.SG
війни — war.GEN.F.SG
і — and
логікою — logic.INS.F.SG
вулиць — streets.GEN.PL

The speaker simultaneously wanted to say that it was necessary to abandon ‘dictated methods of political struggle’ and ‘actions dictated by the logic of the streets’. It all came out together.

[dan-zeman]

Your tree looks good to me.

[shvedova]

Thank you, Dan. Is there any way to mark that this is not the sentence to learn on?
Maybe it would be nice to mark contamination, it's not the only case I've seen. Once I tagged a token with an incorrect grammatical form with a Typo tag, but in fact, it was not just a typo, there was a contamination. I think it can happen regularly in spoken corpora if the recordings are not edited.

[amir-zeldes]

this is not the sentence to learn on?

If this is the correct tree then why is it bad to learn from it? If a parser were to encounter such a sentence in the wild, wouldn't you want it to behave they way the correct tree does?

[shvedova]

There are no such regular sentences in Ukrainian. These are two trees that have fused together in an abnormal way. This speaker was just nervous because they were cancelling the results of the presidential election that day.
It looks like this: "The decision that the government made and the protests on the streets show the urgency of reform was long overdue."

[amir-zeldes]

Yes, I understand, but I think it's legitimate - people get nervous and say weird and ungrammatical things. In terms of learning from this sentence, the tree assigned to it is a clue for how parsers should behave the next time they are confronted by a broken sentence like this. I'm not saying a treebank should strive to over-represent cases like this, but if this is what we would like to happen whenever a parser sees something like this, and it occurred naturally, then I see no reason to exclude the sentence.

[shvedova]

I don't mean to exclude it, it's interesting. My question is whether there is any way to tag it specifically.

[amir-zeldes]

If you know what the person is trying to say, you could provide a target hypothesis annotation indicating the correct form. Otherwise you could consider annotating erroneous sentences using an error tag without a canonical form.

[dan-zeman]

We have guidelines for annotating typos in individual words but we don't have guidelines to annotate errors in syntax. Maybe we should propose something. It would have to be a MISC attribute (at least until we have UD v3 with some implementation of Joakim's proposal of relation features).

[sylvainkahane]

Sorry, I am not very familiar with Ukrainian. Can you explain where is what you call an error of syntax? Is it the fact "і логікою вулиць" 'and logic.INS.F.SG streets.GEN.PL" is not in the "right" place? In this case it is likely to be an afterthought, which is very common in spoken data. We can just introduce a feature such as Afterthought=Yes.

[ellepannitto]

I also am not familiar with Ukranian, so I don't clearly see the issue...
Generally speaking I don't think it's a good idea to mark something as "not to learn on", otherwise we'd be implicitly giving a grammaticality judgement and grammaticality itself is a highly problematic category.
If the text had some issues because, I don't know, it was badly transcribed, it's broken for some external reason or it's just not meaningful, then I'd exclude it from the treebank. But this seems not to be the case here: if this is a strand of spoken data that someone has uttered at some point, it is a legitimate part of language and needs to be described.

From your analysis it seems a somehow complex case of re-planning. What would be the ideal analysis for you?
I mean, if you had no constraints on the fact that the representation should be a well formed tree etc...?

If you can analyze it as a re-planning, I would choose which of the two ("actions dictated" or "dictated methods") is the false start and which of the two is the repair, then I would fully analyze the repair and treat the rest as a false start, so that if you filter out the false start subtree the sentence seems acceptable to you.

[bulbulistan]

@sylvainkahane The issue is rather that - according to @shvedova's explanation, congruent with my understanding - "надиктованих методів політичної війни" is inserted at an incorrect place, so дій ends up dangling there, where the speaker originally wanted to modify it by "надиктованих логікою вулиць".
The analysis suggested by @ellepannitto seems very reasonable, I would vote for it.